Unsupervised representation learning for time series anomaly detection and beyond

Abstract

As industrial processes are digitised, growing volumes of time series data must be monitored for anomalies. Set in the context of automotive powertrains, this thesis investigates truly unsupervised anomaly detection: methods requiring no labelled data at any stage, including for setting the detection threshold.

Progress in this field is hindered by a lack of high-quality public benchmarks and by loose terminology, as many supposedly unsupervised methods still rely on labels for threshold calibration. This thesis contributes PATH, a publicly available multivariate time series dataset built from physically motivated simulation models, combining real-world-like complexity with precise control over anomaly generation and ground-truth labelling.

It then benchmarks a novel unsupervised method against the state of the art, showing that the unsupervised threshold is the weakest link, with performance dropping sharply when training data is contaminated with anomalies. To improve threshold choice, a novel dissimilarity-based active learning query strategy is proposed, maximising the diversity of sequences labelled by a domain expert. It outperforms traditional query strategies, especially at low query budgets, and remains robust even when many queries are mislabelled. Finally, the same modelling procedure is applied to predictive maintenance, detecting deterioration on a real-world test bench well before an existing rule-based monitor.